Agentic commerce is the shift from AI that only recommends a product to AI that moves the transaction forward. It checks live availability, builds a cart, applies the right price, and completes checkout inside limits a business sets in advance. OpenAI’s current checkout flow requires the shopper to confirm each step. Google’s Agent Payments Protocol goes further. It also supports delegated purchases, where an agent can act without a new approval once the user has pre-authorized the merchant, spending limit, timing, and other conditions. Both are still a meaningful shift from a recommendation widget. eMarketer expects AI platforms to drive $20.9 billion in US retail spending in 2026, nearly four times what they drove in 2025.
If you run a retail platform, a commerce backend, or anything that touches checkout, agentic commerce has already arrived, even with a human still confirming most orders. The real question is narrower. Can your systems hand a purchase decision to software you don’t control, verify what that software was authorized to do, and still trust what comes out the other end?
From recommendation engines to agentic commerce
Retail AI used to stop at the suggestion. A recommendation widget surfaced a product, a chatbot answered a question, and a person decided what happened next. Agentic commerce removes that last step, at least in delegated scenarios. Once a shopper has pre-approved a merchant, a spending limit, and a timeframe, the agent doesn’t need a new approval for every routine purchase. It compares prices across sellers, checks the approved budget, and completes the transaction within those limits.
Adobe’s most recent traffic data shows why this stopped being a niche behavior. AI-referred traffic to US retail sites rose 393% year over year in the first quarter of 2026, TechCrunch reported. That traffic converted 42% better than regular traffic in March 2026, a full reversal from March 2025, when AI traffic converted 38% worse. Those AI-referred shoppers also spent 48% more time on site and generated 37% more revenue per visit. These visitors tend to arrive further along in the decision process than shoppers from traditional channels. That’s still AI-assisted discovery rather than fully delegated purchasing. But it shows a meaningful part of the shopping journey has already moved into AI interfaces, and retailers are starting to build for that reality instead of treating it as a dashboard curiosity.
Salesforce just made it official
On July 6, 2026, Salesforce took Agentforce Commerce out of pilot and into general availability. Treat that date as a signal more than a product update. The Shopper Agent now carries a customer from discovery through checkout on a retailer’s own storefront. The Buyer Agent handles B2B orders over WhatsApp and SMS without a portal login. The Merchant Agent runs back-office catalog and promotion work in plain language instead of a rules engine.
The distribution move matters more than the feature list. Salesforce confirmed native integration into ChatGPT this same month. Google Search’s AI Mode and the Gemini app follow later in the summer. That means a shopper’s AI agent, running in an interface a retailer doesn’t own, can complete a purchase on that retailer’s storefront directly. Salesforce has reason to move fast here. According to its own holiday season data, retailers running their own shopper agents grew sales 59% faster than retailers that sat out. AI-referred traffic converted at roughly eight times the rate of social traffic. When a platform this size ships this fast, “watch this space” stops being the right response.
Two sides of agentic commerce: buying and retail operations
Salesforce’s own product split is a useful map here. Shopper Agent works the customer-facing side: search, cart, checkout. Buyer Agent handles B2B ordering. Merchant Agent runs catalog and promotion work behind the scenes. Those are two different problems. Customer-facing agentic commerce is about a shopper’s agent finding and buying a product, still with a confirmation step in most flows today. Agentic retail operations is about a retailer’s own systems making decisions autonomously, on pricing, inventory, and storefront content, without a customer in the loop at all. Three patterns matter most on that operational side: dynamic pricing, autonomous inventory decisions, and session-level storefront personalization.
Dynamic pricing is the clearest case. A traditional pricing system runs on fixed rules: match a competitor minus 5%, hold a margin floor, open a promotional window on schedule. An agentic pricing system continuously optimizes price within predefined margin, compliance, and promotion guardrails, using live competitor, demand, and inventory signals instead of a static rule set.
Inventory management is where the dollar figures get concrete, and it’s the pattern with the clearest production evidence. Walmart reports that its Self-Healing Inventory system automatically reroutes overstock to the stores that need it, before the surplus becomes a write-off. It has saved the company more than $55 million so far. Walmart is now extending the system beyond the US into Costa Rica, Mexico, and Canada.
Personalized storefronts are the least visible pattern, but arguably the highest-leverage one where it’s deployed. An agentic system doesn’t work from pre-built customer segments. It rebuilds page layout, product order, and promotional content for each session in real time, and no marketer approves each configuration individually.
The agentic commerce architecture problem hiding underneath
Most omnichannel builds solved a coordination problem: keep the cart synced between mobile and desktop, match an in-store promotion to the website. That got fixed at the interface layer. The backend systems underneath, point of sale, ERP, warehouse management, CRM, mostly stayed separate. They only needed to look consistent to a human shopper. Nobody required them to talk to each other in real time.
Agentic commerce breaks that assumption. An agent making a pricing call needs live inventory data. An agent rebuilding a storefront needs current margin data by SKU. When those systems don’t share a common data layer, the agent acts on stale or incomplete information. A wrong decision made instantly is often worse than a slow one.
Unified commerce is the usual architectural answer. It’s a shared, real-time operational layer that gives every channel a consistent view of products, pricing, inventory, orders, and customer data. That doesn’t always require replacing every backend system, but it does require removing the conflicting versions of truth between them. It’s the prerequisite most agentic commerce vendor demos skip past. Manhattan Associates’ 2026 benchmark covered more than 400 specialty retailers across North America, EMEA, and Latin America. Only 7% qualified as unified commerce leaders, while 33% remained in the Basic category. The leaders posted nearly twice the growth rate of the least mature retailers. That gap is the real bottleneck, not the AI model choice.
What’s actually blocking most retailers from agentic commerce
The gap isn’t primarily budget. Most retailers are blocked by infrastructure decisions made five to ten years ago, decisions that were reasonable at the time.
Legacy point-of-sale systems are the most common chokepoint. They were built to record a transaction, not to feed a real-time data layer. Pulling live sales data out of a legacy POS without a full replatform usually means middleware, custom connectors, and ongoing maintenance. Every extra layer between the source system and the agent adds latency and another point of failure.
Fragmented data is the second blocker. Customer records sit in a CRM. Inventory sits in an ERP. Web behavior sits in an analytics platform. Margin data sits in a spreadsheet finance updates monthly. Commercetools’ research on agentic commerce readiness found that 81% of retailers say data quality issues affect business decisions at least sometimes, 45% sometimes and 36% often. An agent can’t reason across data that doesn’t connect. Unifying that data has to happen before any agent goes live, not after.
The third blocker gets the least attention: authorization. Who in the organization actually approved an AI system to change a price, or cancel a supplier order? At most companies, nobody has written down what an agent is allowed to do, under what conditions, or who signs off. Without that, even technically capable systems sit in pilot mode for months, because nobody wants to be the person who flips the switch.
A commerce agent needs more than API access
Giving an agent a live data feed is the easy part. The harder part, and the part most agentic commerce coverage skips, is proving afterward that the agent only did what it was supposed to do. That comes down to four things.
Identity: which agent made the request, and on whose behalf. Authorization: what spending limit, merchant list, or category it’s allowed to touch. Execution safeguards: idempotency so a retried request doesn’t double-charge or double-order, spending caps, and a rollback path. Audit: a record of which cart got approved, by what mandate, using what data.
This isn’t theoretical. Google’s Agent Payments Protocol builds a cryptographically verifiable chain between what the shopper asked for, what got approved, and what was actually charged. An Intent Mandate captures the shopping request and its limits. A Cart Mandate records the exact items and price. A Payment Mandate carries that approval to the payment network. Depending on the flow, the shopper approves in real time or authorizes the agent to act later within those predefined limits. Either way, the chain of signed mandates leaves an audit trail: what was requested, what the agent was allowed to buy, and what was paid. OpenAI’s Agentic Commerce Protocol takes a related approach. OpenAI never becomes the merchant of record. Each delegated payment token is capped at a specific amount and expiry, tied to a specific merchant, before the retailer’s own payment processor ever sees the request. Retailers building for agentic commerce need this same kind of scoped, logged authorization layer, not just an open API. That applies whether they’re exposing a storefront to someone else’s shopping agent or running their own pricing and inventory agents.
What building a POS-connected retail platform taught us
We covered the full build of a POS-connected customer engagement platform in an earlier case study, for a US retail startup with dealer and partner networks. The short version: no unified data layer, a POS replatform that was never on the table. We built a connector architecture on .NET, AngularJS, Google Cloud, and Kubernetes instead, surfacing POS data in near real time without touching the source system.
What that project made clear applies directly to agentic commerce. Partial stock updates. Delayed POS syncs. Two channels reporting different numbers at the same time. Those are the edge cases where agentic systems either hold up or fall apart.
Handing that same connector layer to an external AI agent, rather than only to your own dashboards, adds a new list of requirements. A machine-readable product catalog the agent can actually parse. A real-time inventory endpoint instead of a nightly sync. Scoped permissions per agent rather than one shared API key. An idempotent checkout path. A defined fallback for when the POS sync lags. A log of every action an agent took and why.
Where to start if you’re not starting from zero
The retailers making real progress this year didn’t launch a company-wide transformation program. They picked a narrow scope, proved it, and expanded from there.
Start with data unification before touching any agent. Map what data exists, where it lives, and what’s missing. Build or buy the connector layer that makes POS, inventory, and customer data available in one place. Treat that as the slower, unglamorous prerequisite it actually is.
Once that layer is stable, pick the highest-value, lowest-risk use case for a first agent: dynamic pricing in one category, or demand-driven replenishment for the top-selling SKUs. Define its scope narrowly. Build the human override before launch, not after an incident forces one. Run it against a control group for 60 to 90 days, and measure the result honestly before expanding.
Gartner projects that 40% of enterprise applications will include integrated task-specific AI agents by the end of 2026, up from less than 5% in 2025. Bain estimates agentic commerce will make up 15% to 25% of total US ecommerce sales by 2030, a $300 to $500 billion market. McKinsey’s global estimate runs as high as $3 to $5 trillion by the same year.
That’s a lot of money chasing a problem most companies haven’t solved yet. Data and governance work, done early and without fanfare, is what separates the two outcomes. One retailer scales an agent. The other spends 2027 explaining why the pilot never launched.
How Allmatics approaches this
Our Web/Mobile Development and AI/ML Development work covers exactly this layer. That means connector architecture exposing POS, inventory, and customer data in real time without a forced replatform. It also means agent scoping that separates what a system can suggest from what it’s authorized to execute on its own. We wrote up the full POS-connected retail SaaS build if you want the detail behind the architecture referenced above. Our broader view on retail and e-commerce infrastructure is a useful starting point if you’re earlier in scoping this.
If you’re deciding where an AI agent should sit in your retail stack, get that decision right during product discovery. It’s cheaper than fixing it after the agent is already live against production data.
Frequently Asked Questions
What is agentic commerce? Agentic commerce is retail AI that acts on a shopper’s or a business’s behalf: comparing options, adjusting prices, managing inventory, or completing a purchase, rather than only recommending a product for a human to act on. It requires real-time data access and defined authority to execute, not just to suggest.
Is agentic commerce the same as a shopping chatbot? No. A chatbot answers questions and waits for a human decision. An agentic system takes the next step: executing a price change, a reorder, or a checkout, within limits a business defines in advance.
What does a retailer need before deploying an agent? A unified data layer connecting POS, inventory, and customer data in real time, a narrowly scoped first use case, and an explicit governance rule for what the agent can do without human approval. Skipping the data layer is the most common reason pilots stall.